Conversation
|
Thank you for your contribution. I am currently busy with my PhD defense, which may take another week or two. After that, I will do some code review if I have time. :) Hopefully this won't take too long. Best, |
|
No worries, take your time, and good luck with your defense! |
|
Here are the benchmark results on the SimBEV dataset test set for 3D object detection and BEV segmentation, respectively. The BEV segmentation results are somewhat suprising. I was wondering if you have any takes on why UniTR does not perform as well as BEVFusion for the bus class, and why adding LSS improves the results for the larger objects but makes it worse for smaller ones, especially pedestrian. |
|
As for your custom dataset, I wonder if UniTR+LSS has better performance on mIoU in BEV seg mission? |
|
It doesn't seem to be better at mIoU... the results are in the table in my previous post, it improves on UniTR for larger objects but gets worse for smaller ones. |


Hi,
This pull request adds support for the SimBEV dataset. As far as I know the additions have no impact on how the code works for other datasets.